1. Field of the Invention
The present invention relates to biometric identification using an iris image, and more particularly, to identification of the iris image in an eye image for extracting iris data.
2. Description of Related Art
Due to the unique character of each individual's iris, various systems attempt to use the iris for biometric identification. Such systems generally capture an image of the entire eye, which includes an image of the iris. The iris image must be identified before the patterns of the iris, which are unique to each individual, can be extracted for biometric analysis. In other words, the area that corresponds to the iris must be segmented, or separated, from the other components in the entire eye image. Conventional systems generally determine the boundaries of the iris image by searching for the edges that correspond with these boundaries. In particular, these conventional approaches depend on the contrast between the edges and the area around the edges, referred to as edge strength, to identify the boundaries. As described further below, this approach suffers from many disadvantages and fails to provide a robust method for finding and extracting iris data. For instance, because the edges between the iris and the sclera (limbic boundary) are often weak and hard to detect, conventional systems are beset with the difficult challenge of enhancing the weak edges to identify the iris boundaries adequately.
U.S. Pat. No. 5,291,560 to Daugman implements an integro-differential operator for locating the circular edges of the iris and pupil regions, as well as edges of the arcs of the upper and lower eyelids. As a first step, Daugman approximates the edge of the pupil to be a circle and sums the brightness along each circle with trial center coordinates (x0, y0) and incrementally increasing radius r. When radius r reaches the edge of the pupil, there is a sudden change in the brightness since the brightness of pixels in the pupil region should be different from pixels in the iris region just outside the pupil region. Various trial center coordinates (x0, y0) can be evaluated to find the center of the pupil. If the center coordinates (x0, y0) do not coincide with the pupil's center, some portions of the circle will still lie within the pupil region when the brightness suddenly changes. However, if the center coordinates (x0, y0) do indeed coincide with the pupil's center, when the brightness suddenly changes, no portions of the circle should be in the pupil region, so the rate-of-change of the brightness, or luminance, should be at its maximum. Thus, the problem of locating the pupil's boundary is reduced to an optimization problem where a three-parameter space is searched for the best combination of center coordinates (x0, y0) and radius r, i.e., where the absolute value of the partial derivative with respect to radius r of the integrated luminance along the circle is maximum. The search of the three-parameter space can occur in an iterative process of gradient-ascent.
As a second step, Daugman's approach for finding the outer edge of the iris, also known as the limbic boundary, is similar to finding the edge of the pupil but the previous approach is modified to account for the fact that i) the pupil is not always centered in the iris, ii) the upper and lower eyelids obscure top and bottom portions of the iris, and iii) the iris, unlike the pupil, has a concentric texture and may itself contain interior circular edges which could create sudden changes in integrated luminance along a circle. The process for detecting the iris edge therefore is restricted to two 45-degree arcs along the horizontal meridian and an area integral is used rather than a contour integral. The luminance for arcs of increasing radius and centered at the pupil center are evaluated as an area integral in polar coordinates. The value of radius r which corresponds to the maximum in the rate-of-change of integrated luminance with respect to radius r corresponds to an edge of the iris. This calculation is made for each arc separately since the left and right edges of the iris may be at different distances, i.e. radius r, from the pupil's center.
Disadvantageously, as Daugman acknowledges by using area integrals when processing the iris image, the algorithm can fail because it is susceptible to changes in luminance that do not occur at the boundaries, such as those caused by reflections, noise from a poor image, contact lens edges, or even by actual features or textures in the eye. In particular, the use of integro-differential operators are sensitive to the specular spot reflection of non-diffused artificial light that can occur inside the pupil, and such spots can cause the detection of the correct inner boundary to fail. Therefore, Daugman has also proposed the use of Gaussian filtering to smooth the texture patterns inside the iris region to avoid incorrect detection of false limbic boundaries, but this approach involves heavy computational complexity.
In addition, the process above must be modified to account for the fact that the edges are not clean edges and are somewhat fuzzy. The integrated luminance of a shell sum must be used rather than the integrated luminance of the circle, i.e. the rate-of-change of a shell sum is maximized. Therefore, use of the Daugman may require ad hoc adjustments of the shell size parameter.
U.S. Pat. Nos. 5,751,836 and 5,752,596 to Wildes et al. also implement edge detection algorithms. The process disclosed by Wildes et al. initially averages and reduces the input image using a low-pass Gaussian filter that spatially averages and reduces high frequency noise. The result is then subsampled without further loss of information but with the advantage of reducing computational demands. The iris is then localized by locating the limbic (outer) boundary of the iris, the pupillary (inner) boundary of the iris, and the eyelid boundaries. The iris is then taken as the portion of the image that is outside the pupillary boundary, inside the limbic boundary, above the lower eyelid, and below the upper eyelid.
The first step in locating each component of the iris boundary employs a gradient-based edge detection operation which forms an edge map by calculating the first derivatives of intensity values and then thresholding the result. Wildes et al. bias the derivatives in the horizontal direction for detecting the eyelids, and in the vertical direction for detecting the limbic boundary. The application of the process taught by Wildes et al. has the disadvantage of requiring threshold values to be chosen for edge detection. Poor choice of threshold values can eliminate critical edge points possibly causing failure in the detection of the circles and arcs making up the boundaries of the iris.
Using the resulting edge map, the second step employs a transform like that generally disclosed in U.S. Pat. No. 3,069,654 to Hough. The limbic boundary is modeled as a circle with center coordinates (x0, y0) and radius r. Thus, the detected edge pixels from the edge map are thinned to increase the number of meaningful edges. The pixels are then histogrammed into a three-dimensional space formed by circle parameters x0, y0, and r (Hough circle transform). The (x0, y0, r) point with the most number of votes from the histogramming process then represents the limbic boundary. Similarly, the pupil is also modeled as a circle and the edge pixels are thinned and histogrammed into (x0, y0, r) values, where the (x0, y0, r) point with the most votes are taken to represent the pupillary boundary. The eyelid boundaries, however, are modeled as two separate parabolic arcs. The eyelid edges are thinned and histogrammed according to the parameters necessary to define a parabolic arc (parabolic Hough transform), where the set of parameters with the most votes is taken to represent the upper or lower eyelids.
In the article titled “Recognition of Human Iris Patterns for Biometric Identification” by Libor Masek, the Hough circle transform is also used, but unlike Wildes et al., Masek first employs Canny edge detection to create the edge map. Masek modifies Kovesi's Canny edge detection to allow for the weighting of gradients as Wildes et al. teaches for the detection of the limbic boundary with a vertical bias. The Hough circle transform is applied to find the limbic boundary first and then the pupillary boundary where the range of radius values for the search is set manually depending on the database used and the typical radius values in the images.
With respect to the eyelids, Masek applies Canny edge detection where only horizontal gradient information is taken. The eyelids are then located by fitting a line to the upper and lower eyelids through a linear Hough transform. A second horizontal line is formed from the intersection of the fitted line and the limbic boundary closest to the pupil in order to achieve maximum isolation of the eyelids. The use of the linear Hough transform is less computationally demanding than the use of the parabolic Hough transforms taught by Wildes et al. However, maximum isolation of the eyelid regions can isolate substantial portions of the iris itself and make the matching process less accurate.
Disadvantageously, Masek requires threshold values to be specified to create the edge maps, and as with Wildes et al., these threshold values are dependent on the database and the quality of images in the database. Poor choice of threshold values can eliminate critical edge points possibly causing failure in the detection of the circles and arcs making up the boundaries of the iris.
In “Experiments with an Improved Iris Segmentation Algorithm” (Fourth IEEE Workshop on Automatic Identification Advanced Technologies (AutoID), October 2005, New York), Lui et al. disclose a process known as ND_IRIS, which attempts to improve Masek's approach toward segmentation. Masek's method detects the outer iris boundary first and then detects the inner boundary within the outer boundary, but Lui et al. reverses this order by detecting the inner boundary first since there is often greater contrast between the pupil and iris, and thus the inner boundary can be easier to localize. Lui at al. also point out that edge pixels that do not lie on the iris boundary often cause the Hough transform to find an incorrect boundary. Thus, edges within the pupil and the iris are further reduced through thresholding before Canny edge detection. In addition, Lui et al. modifies the Hough transform and implements a verification step which compares the areas on both sides of a proposed boundary. Also, each eyelid is modeled as two straight lines rather than just one. Despite improved iris segmentation, ND_IRIS, like the Masek approach, fails particularly when the image quality is low. Published results also suggest that ND_IRIS problematically has higher recognition rates for light iris images than for dark iris images.
The conventional processes above are not robust because they all require searching for parameters that define the shapes that approximate the iris boundaries. Daugman searches for the best combination of center coordinates (x0, y0) and radius r to define a pupil circle and arcs that mark the left and right edges of the iris. Meanwhile, Wildes et al. applies the circle and parabolic Hough transforms which creates a histogram in to a parametric-space, e.g. a three-dimensional (x0, y0, r) space for the circle Hough transform, to define circular and parabolic shaped boundaries of the iris. Similarly, Masek applies circle and linear Hough transforms.
In addition, approaches that rely on edge strength to detect eye components may require the use of a edge strength threshold. Generally, the edge strength threshold must be determined adaptively at a high computational cost. If the selected edge strength is too high, only the edges from the eyelids and eyelashes can be identified. On the other hand, if the selected edge, strength is too low, too many edges are identified for analysis. As a result, a fourth parameter corresponding to the threshold for edge strength, in addition to x, y positions and radii, must be searched. The edge strength threshold is determined iteratively by processing the database of captured images. If the camera settings or lighting conditions change, however, a new edge strength threshold must be recalculated by iteratively processing the database of images captured under the new camera settings or lighting conditions. As such, the need to recalculate parameters every time conditions change makes the conventional edge-detection systems above less robust.
In addition, the processes above can be highly dependent on the quality of the images being analyzed. For instance, Daugman requires a shell size parameter to be adjusted while Wildes et al. and Masek require threshold values to be chosen for edge detection.
Furthermore, because techniques, such as those taught by Daugman and Wildes et al., identify edges by analyzing localized areas of the image in an incremental manner, the techniques are computationally costly if the approximate location of the iris is not known. As such, a full search along the x- and y-axes as well as the radii for the entire image may be required.
In “Efficient iris recognition by characterizing key local variation” (IEEE Transactions on Image Procession, vol. 13, no. 6, 739-50, 2004), Ma et al. attempt to lower the computational cost by using a “global” projection along the x- and y-axes to find the minima position as the approximate position of the iris. A narrower region, e.g. 120×120, centered on the point is then binarized to a reasonable threshold using the grey level histogram twice to find the center. The lighting of the iris, however, usually creates a bright spot of spots inside the pupil, and results in the minima of the xy projection to be at the wrong location.
Despite the disadvantages of the approaches described above, current attempts to improve segmentation continue to focus exclusively on analyzing edges particularly through parametric approaches, namely the methods used by Daugman and Masek. As described above, Lui et al. uses a segmentation approach based on Masek. In addition, in the article “Person identification technique using human iris recognition,” Tisse et al. implement an improved system based on the integro-differential operators as taught by Daugman in combination with a Hough transform. Moreover, in “A Fast Circular Edge Detector for the Iris Region Segmentation” (Biologically Motivated Computer Visio: First IEEE International Workshop, Seoul, Korea, May 15-17, 2000), Park et al. propose using a method based on Daugman where the need for Gaussian filtering proposed by Daugman is eliminated by searching from a radius r that is independent of the texture patterns in the iris to find the limbic boundary.